Personalised Prediction of Self-Reported Emotion Responses to Music Stimuli
In this thesis I develop a robust system and method for predicting individuals’ emotional responses to musical stimuli. Music has a powerful effect on human emotion, however the factors that create this emotional experience are poorly understood. Some of these factors are characteristics of the music itself, for example musical tempo, mode, harmony, and timbre are known to affect people's emotional responses. However, the same piece of music can produce different emotional responses in different people, so the ability to use music to induce emotion also depends on predicting the effect of individual differences. These individual differences might include factors such as people's moods, personalities, culture, and musical background amongst others. While many of the factors that contribute to emotional experience have been examined, it is understood that the research in this domain is far from both a) identifying and understanding the many factors that affect an individual’s emotional response to music, and b) using this understanding of factors to inform the selection of stimuli for emotion induction. This unfortunately results in wide variance in emotion induction results, inability to replicate emotional studies, and the inability to control for variables in research. The approach of this thesis is to therefore model the latent variable contributions to an individual’s emotional experience of music through the application of deep learning and modern recommender system techniques. With each study in this work, I iteratively develop a more reliable and effective system for predicting personalised emotion responses to music, while simultaneously adopting and developing strong and standardised methodology for stimulus selection. The work sees the introduction and validation of a) electronic and loop-based music as reliable stimuli for inducing emotional responses, b) modern recommender systems and deep learning as methods of more reliably predicting individuals' emotion responses, and c) novel understandings of how musical features map to individuals' emotional responses. The culmination of this research is the development of a personalised emotion prediction system that can better predict individuals emotional responses to music, and can select musical stimuli that are better catered to individual difference. This will allow researchers and practitioners to both more reliably and effectively a) select music stimuli for emotion induction, and b) induce and manipulate target emotional responses in individuals.